See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/265217098

Automatic Lineament Detection Using Digital Elevation Models with Second

Derivative Filters

Article in Photogrammetric Engineering and Remote Sensing · April 1999

CITATIONS READS

31 300

1 author:

David Wladis
Chalmers University of Technology
9 PUBLICATIONS 101 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Hydrogeological risk analysis View project

All content following this page was uploaded by David Wladis on 25 December 2015.

The user has requested enhancement of the downloaded file.

 Automatic Lineament Detection Using Digital
Elevation Models with Second Derivative Filters
David Wladis

Abstract across the entire window, with the result that the direction
A grid operator designed for the analysis of potential field of linear features rather than their spatial distribution is de-
data has been applied to a digital elevation model for the tected.
detection of lineaments. The results obtained have been eval- This paper suggests that a second vertical derivative fil-
uated using both an existing lineament map of the area and tering technique, in which the filter weights are optimized
a field investigation. The evaluation indicates that the sug- for the study region, be applied to digital elevation models
gested methodology facilitates lineament detection. The high (DEMS) for lineament detection. The digital filters employed
degree of consistency between the suggested method and the were developed for analyses of potential field data such as
existing tectonic map indicates that lineaments in the study magnetometry and gravimetry. By constructing second verti-
area are largely topographic features. Field validation con- cal derivative maps, qualitative and quantitative interpreta-
firmed these findings except i n a few well-defined situations. tions can be made more objective in the example of magnetic
maps (Henderson and Zietz, 1967). The basic concept is that
the data are convolved with a mathematical function in
Introduction terms of a set of weights, which can be interpreted as a filter.
In geology and hydrogeology, remote sensing is often used The appropriate set of weights should be selected on the ba-
for the detection of brittle tectonic structures such as faults, sis of its two-dimensional frequency response, which, in
large-scale fractures, and fracture zones. These structures are turn, may be data-dependent as discussed later.
commonly denoted as lineaments (O'Leary et al., 1976). Both In the study area, thorough lineament mapping was car-
satellite images and aerial photographs as well as other types ried out by the Swedish Geological Survey (SGU)and pre-
of remote sensing data are used to detect lineaments. The sented as a tectonic map (Samuelsson, 1978). The mapping
technique can be used in applications such as groundwater was done from aerial photographs at two scales. The sGu tec-
exploration, prospecting for mineral and petroleum re- tonic map facilitated the evaluation of the results obtained
sources, hard rock engineering, and mapping of possible from the automatic detection, with respect to both spatial
pathways for the transportation of pollutants. and directional distributions. A field study was also carried
In a lineament mapping procedure, the lineaments are out, aimed primarily at examining the largest discrepancies
often visually identified by an interpreter (e.g., a geologist), between the automatically detected lineaments and the SGu
which implies that the obtained lineaments are in some way tectonic map.
subjective interpretations and that they often are extracted
manually, e.g., lineaments from the remote sensing data are Digital Elevation Model
hand-drawn on transparent overlays. This methodology pro- The DEM used here was produced by the National Land Sur-
duces results that to a large extent cannot be reproduced vey (LMV). Three main sources for obtaining elevation data
because the identification criteria are not agreed upon by dif- have been employed: aerial photographs, contour lines, and
ferent analysts and usually cannot be expressed in quantita- topographic profiles. Elevation data are used to interpolate
tive terms but, rather, are based on sensory impressions. elevation values on a regular grid with a spacing of 50 m.
Practically all geologic work includes some amount of sub- Because elevations in the DEM consist of interpolated values,
jective interpretation, but it is desirable to minimize this the geometrical accuracy is less than that of directly mea-
kind of uncertainty. Lineament extraction could be more sured points. The error is dependent on the interpolation al-
highly valued if the results were reproducible. This would be gorithm and the topography of the specific area. However,
achieved by using some form of automatic, or criteria-based, the standard error of grid intersection points has been deter-
lineament extraction algorithm. mined to be approximately + 2 m (Ottoson, 1978). For fur-
The idea of using different automated, or criteria-based, ther details on the establishment of the LMV DEM, refer to
lineament-extraction algorithms is not new (Wang and Ho- Ottoson (1978). The original DEM is shown in Figure 1.
warth, 1990; Moore and Waltz, 1983); however, several In this context, the influence of the methodology used to
problems have arisen in the past. Slow algorithms and exten- produce the DEM (i.e., the interpolation algorithm) on the in-
sive manual editing have limited the use of such algorithms. terpretability of lineaments should be considered. The accu-
Other approaches developed for optical data, such as aerial racy of the interpolation of the grid nodes is affected by the
photographs and satellite images, often yield false lineaments relative relief of the area (Ottoson, pers. comm., 1996). How-
that are related to roads, power lines, and other man-made ever, the relative relief of the studied area is not considered
features which are time consuming to correct (Gustafsson,
1994).
The Hough transform is frequently used to detect linear Photogrammetric Engineering & Remote Sensing,
features in images (e.g., Karnieli et a]., 1993). The transform Vol. 65, No. 4, April 1999, pp. 453-458.
is computer intensive, and detected lines are extrapolated
0099-1112/98/6504-453$3.00/0
Department of Geology, Chalmers University of Technology, O 1999 American Society for Photogrammetry
S-412 96 Gothenburg, Sweden ([email protected]). and Remote Sensing

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING April 1999 453

 where KTrepresents the average value of AT (total intensity)
on the circle of radius r, J, is a Bessel function of zero order,
A, are coefficients to be determined, and pkare positive roots
of JO(w).
In order to compute the second vertical derivative of the
total intensity (AT) at the origin, the following equation is
provided:

where the coefficients A, can be computed by solving K

equations in K unknowns, which involves the value of the
anomaly at the center and the average values of AT on K - 1
circles about the origin. This is illustrated in Figure 2.
From Equation 1, the following simultaneous equations
are used to solve for the coefficients in order to compute the
filter for a nine-point convolution matrix consisting of AT,
and eight neighboring points:
AT, = A,
- + A, + A,
dT, = A,Jo(~.c,rl)+ AzJo(~zr1)+ A3Jo(~3ri)
+ A3J0(p3rZ)
AT2 = A,J,(P~~,)+ AZJO(p2r2) (3)
Then r, = S and r2 = ~ (wheref i S is the grid spacing),
if the averaginn circles are transformed to cartesian coordi-
nates (Full&, 1967) as shown in Figure 2. To solve for the
coefficients A,, A,, A, in Equations 3, we take 10s as the ra-
Figure 1.Original National Land Survey (LMV) DEM. dius at which the idealized anomaly is arbitrarily smoothed
to zero, i.e., where AT is essentially zero (Henderson and
Zietz, 1967). Equation 2 can then be written as
extreme compared to other parts of the country, and the ac-
curacy of the DEM is considered to be high.
All image processing of the DEM is carried out using a
standard image processing system, ERMapper 5.0. Subse-
quent analyses were carried out in a geographical informa-
tion system, Genamap 4.2.
which is the working formula. In this study, different sizes of
filters have been computed and tested on the DEM. The grid
operators for different filter sizes have been computed using
1
Matlab 4.2a.
filtering Technique The degree of detail in the final result is achieved
Elevations and other potential field data (e.g., magnetic mainly by modifying the threshold value, which is described
maps) are often represented on a regular grid, so equations
and formulas developed for the analysis of potential field
data are readily applied to DEMs.
S
Different kinds of grid operators are designed to accom-
plish specific purposes on potential field data (e.g., to re- *
move trends or to study features of small areal extent). These
grid operators may be interpreted as filters with frequency
responses in a two-dimensional frequency domain (Fuller,
1967). A variety of methods have been developed for use in
analyses of gravitational and magnetic data, but, for the pur-
poses of this study, a derivative method has been applied.
The theory and design of the grid operators used in this
method are described in BAth (1974) and Fuller (1967). In
this study, the grid operators chosen were the second vertical
derivative operators described by Henderson and Zietz
(1967).
Lineaments in a digital elevation model are defined by a
drop in elevation for a short distance. Therefore, a lineament
can be described by a certain frequency in terms of its repre-
sentation in a digital elevation model. Consequently, we are
interested in the frequency domain of the sampled data and,
thus, the frequency response of the filter.
The essential equations in Henderson and Zietz (1967)
for the computation of filter constants are described next.
The equations are derived in Henderson and Zietz (1967),
who provide the following series expansion:
Figure 2. Graphical relation between averaging rings,
AT(rj, and Cartesian coordinate points, AT.

454 April 1999 PHOTOGRAMMETRICENGINEERING & REMOTE SENSING

 the structures range from narrow fracture zones to broad val-
leys, the applied threshold value may be crucial. By adjust-
ing the threshold properly, classifkation of the structures
based on their prominence is also possible. Classification of
lineaments is performed manually in the compilation of tec-
tonic maps (Ottoson, 1978).
In the filtered image, topographic lows occur as negative
pixel values and topographic highs as positive pixel values,
i.e., the extremities of the slopes in the original DEM are
picked out by the filter. The magnitude of the value is pro-
portional to the gradient of the slope. Lineaments are repre-
sented by some specific level of the negative pixel values.
Therefore, it is convenient to threshold the image. Generally,
the threshold can be set using either of the following meth-
ods. Using existing data, the threshold value can be cali-
brated from a few known structures in the area or from an
existing tectonic map. Unfortunately, such data are not al-
ways available. The threshold may also be set using profes-
sional judgement. To some extent, this introduces subjectiv-
ity into the mapping process, but the mapping of structures
is consistent throughout the area and the results are repro-
ducible.
Thresholding of the image also provides line thinning. In
this instance, pixels 5 - 5 were set to 255 (gray in Figure 4)
and all other pixels to 0 (white in Figure 4). The resulting
Figure 3. National Land Survey (LMV) DEM filtered using image is shown in Figure 4. Single pixels and minor clusters
second vertical derivative filter. of pixels still remain in the image. These pixels represent
high frequency features of small areal extent such as small
lakes and pits.

in the following section. The detection limit of structures Validation of Results

(i.e., the cut-off limit) is governed not by the filter size but by The most time-consuming aspect of studies comprising ex-
the characteristics of the filter function. For larger filter sizes, traction of lineaments is evaluating and validating results in
the detection frequency is better d e h e d . The filter yields the field. Fortunately, a comparison of the results obtained
values of sufficient accuracy for a 3 by 3 filter matrix in the with an existing tectonic map was possible. Furthermore, in
geomagnetic example described by Henderson and Zietz this restricted pilot study, a thorough field examination was
(1967). In order to study and compare the effect of different considered to be an important component.
grid operators, frequency analysis can be used (Fuller, 1967). A tectonic map of the study area was published by SGU
The grid spacing may also influence the results. A h e (Samuelsson, 1978), where lineaments were interpreted from
grid spacing produces more accurate results in terms of de- aerial photographs of two different scales, 1:33,000 and
gree of detail. However, for certain purposes, only the re- 1:65,000. A systematic control in the field of the lineaments
gionally developed lineaments are of interest. By increasing obtained has not been undertaken by SGU.However, the de-
the grid spacing in proportion to the size of the structures to gree of error of the mapped lineaments is empirically esti-
be studied, minor and local structures can be removed (Hen- mated by Samulesson (1978) at less than 1 percent (i.e.,
derson and Zietz, 1967). fewer than 1 percent of the mapped structures lack field evi-
The size of the filter matrix is proportional to the wave- dence). The results of Samuelsson (1978) were compared to
lengths allowed through the filter (Mather, 1987). Therefore, those obtained from the DEM.
broader structures are detected using a larger filter size, The lineaments extracted from the DEM provide a useful
given that they yield a sufficiently high frequency. A filter of map for different geological applications (Figure 4) but, for
this type produces a margin around the image proportional the purposes of this study, they were vectorized by on-screen
to the size of the filter matrix. This margin contains no data. digitizing using two conditions: (1)a lineament consists of at
For a 5 by 5 filter, the margin is two pixels whereas, for a least four pixels (200 m) and (2) there cannot be a gap of
21 by 21 filter, the margin is ten pixels broad. Therefore, a more than one pixel (50 m) between any two groups of at
smaller filter may be preferable. It is worth noting that larger least four pixels, criteria that can easily be automated. The
filters are more computer intensive. map containing the digitized lineaments is shown in Figure
A listing of how the grid operators were computed in 5. It should, however, be noted that the lineaments are digi-
Matlab 4.2a is provided in the appendix. The listing is in the tized mainly to facilitate the preparation of lineament statis-
form of a Matlab script file, including the statements and tics and that the digitizing procedure may lead to at least
data necessary to compute a 5 by 5 filter. The script can be some loss of information.
edited in order to compute an arbitrary filter size. The com- The lineaments obtained from the DEM and lineaments
puted grid operators are used to convolve the DEM. This op- from the SGU map were plotted in separate histograms (Fig-
eration is implemented in all raster-based image processing ure 6). From the histograms, it can be seen that the north-
systems, e.g., ERMapper. The filtered output is shown in and northeast-trending lineament directions are dominant in
Figure 3. both extraction procedures.
The lineaments from the sGu tectonic map and the auto-
Threshold matically extracted lineaments are summarized in Tables 1
The thresholding procedure is applied to extract the relevant and 2, respectively. The lineaments are sorted in 5-degree in-
information from the filtered output. The threshold value can tervals. From the tables, it is clear that the three largest
be set to obtain the desired degree of detail. In areas where peaks determined using each procedure coincide except for a

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING April 1999 455

 \

-
\
Figure 5. The digitized, automatically extracted linea-
0 1 2 ments from the DEM.
Kilometers

Figure 4. Automatically extracted lineaments in raster for-

mat (gray) overlaid by SGU lineaments (black lines).

W N E W N E
2000fJ 20000

15000 15000

a
10000

5000 5000

0 0
0 0 0 0 0 0
e ~ k 9 9 9 8 9
oeoz 88SSI8R38 @ 8 ~ 8 $ 9 @ 5 ~ 0 2 R 8 ~ S 1 8 R 8 a
kmuth Azlrnuth

Figure 6. Lineaments extracted by SGU (left), and from DEM. Arrows indicate three largest peaks.

Min. Max. Mean Total Mean Min. Max. Mean Total Mean
Azimuth Frequency Length Length Length Length Angle Azimuth Frequency Length Length Length Length Angle
NS 11 258.1 3307.5 886.5 9751.1 N 0.3OE NS 37 175.3 3193.2 494.0 18278.5 N 0.4OW
N 40" E 7 376.9 2447.3 1196.8 8377.9 N 40.7" E N 45' E 6 241.4 1272.0 513.1 3078.6 N 43.8OE
N 15" W 9 248.4 1347.8 667.0 6002.7 N 14.3OW N 15" W 2 470.0 1856.9 1163.4 2326.9 N 14.6" W
N 45" E 2 1568.0 3164.7 2366.3 4732.6 N 44.9' E N 5' E 3 462.6 942.4 733.2 2199.6 N 3.8" E

small discrepancy, i.e., the second-largest peak is in the that the mean discrepancy is only 3 degrees. The peak at
N40°E interval in the SGU lineaments, whereas it is in the N15'W in Table 2 (DEM data) consists of only two linea-
N45"E interval in the DEM lineaments. But the tables show ments.

456 April 1999 PHOTOGRAMMETRICENGINEERING & REMOTE SENSING

Field Evaluation Mather, P.M., 1987. Computer Processing of Remotely-Sensed Zm-
ages, John Wiley & Sons, Chichester, 352 p.
The field examination was aimed mainly at evaluating differ-
ences between the two mapping methods. In the lineaments Moore, G.K., and F.A. Waltz, 1983. Objective procedures for linea-
ment enhancement and extraction, Photogrammetric Engineering
mapped by SGU, lineaments were systematically interpolated 8 Remote Sensing, 49:641-647.
and extrapolated. This is particularly clear in paved and land-
O'Leary, D.W., J.D. Friedman, and H.A. Pohn, 1976. Lineaments, lin-
filled areas, where no trace of lineaments could be found in ear, lineations: Some proposed new standards for old terms,
the field. Furthermore, lineaments shorter than approximately Geol. Soc. Am. Bull., 87:1463-1469.
250 m were not recognized by SGU. Lineaments represented Ottoson, L., 1978. Establishment of a high density terrain elevation
by a "step" or a broad valley tend to be under-represented in data base in Sweden, Proc. 9th International Conference on Car-
the automatic mapping procedure. Under-representation of tography, International Cartographic ~ s s o c i a t i o n , - ~ a r ~ l ain
nd,
narrow structures is possibly c o ~ e c t e dto the resolution of National Land Survev (LantmateriverketlProfessional Pauers.
the DEM, which is about ten times less than that of the aerial Giivle, Sweden, 15 p: '
photographs used by SGU. A prominent, slightly curvilinear Samuelsson, L., 1978. Description of the Map of Solid Rocks Gdte-
lineament mapped by SGU (indicated by an arrow in Figure borg SO, Sveriges Geologiska Undersijkning, sene Af, no. 117,
4) was not mapped in the automatic procedure. map scale: 1:50,000, Stockholm, Sweden, 85 p.
i,
Wang, and P.J. Howarth, 1990. Use of the Hough transform in au-
Conclusions tomated lineament detection, LEEE Trans. on Geoscience and
The lineaments detected coincide largely with the linea- Remote Sensing, 28(4):561-566.
ments mapped by s G u in terms of strike directions. Discrep- (Received 15 September 1997; accepted 09 April 1998; revised 25
ancies detected in the field favor the automatically detected May 1998)
lineaments except for two situations, i.e., when the mapped
lineaments are slightly curvilinear or when they are repre-
sented by "steps" in the terrain. Therefore, it can be con-
cluded that the suggested technique extracts appropriate
lineaments from DEMs. The resolution of the DEM and the be-
havior of the algorithm when applied to curvilinear linea- Appendlx
ments should be considered. The results also suggest that
lineaments in this area, to a large extent, are represented by %%%% This is the Matlab scdpt ffle %%%%
topography.
k
A threshold v ue can be applied to produce the most
suitable output for the intended use. The output map may
%% positive mots to Jo(p& 46%

mot=[2.4048255577 5.5200781103 8.6537279129 11.7915344391 14.9309177086

range from only the largest structures of regional impact to
all structures, including the smallest structures inherent in 18.071063%79]
the original data. Furthermore, the criteria applied in the dig-
56% Circle mdius, r %%
itizing procedure is arbitrary. As pointed out above, the ras-
ter-to-vector data conversion may cause a reduction in the radii=[O 1 sqrt(2) 2 sqrt(5) sqrt(8)l
information content in the final map. Therefore, the chosen
criteria for digitizing is dependent on the intended applica- %% compute Jo(pf,) lsrd put them in a m a %%
tion. In this study, results are based exclusively on elevation
data. Integrating other data such as satellite data and aerial for i=1:6;
photographs, if available, may further improve results. for j=l:6;

Acknowledgments rnat(ij)=besselj(O,root(i)*r~ii(i)IlO);
The Swedish National Space Board is gratefully acknowl- end
edged for the funding of this project. Thanks are due to Pro-
fessor Mike Middleton and Erik Larsson at the Department of
Geology, Chalmers University of Technology, for guidance at
various stages of the work. 76% SOIVC tke ~irnultcrnc~~s
~q.ucrti01~
(3) 46%

A=inv(mat)*(mot.*mot)'1100

References 9%-form A, to ccntes~ coordimes %%

B&th,M., 1974. Spectral Analysis in Geophysics: Developments in b = ~ . * [ l114 114 114 118 1/41'
Solid Earth Geophysics 7, Elsevier Scientific Publ. Co., Amster-
dam, 563 p. 46% Tke grid opcrrrton orgcmised in a convolution m a 56%
Fuller, B.D., 1967. Two-dimensional frequency analysis and design c(3,3)=b(l);
of grid operators, Mining Geophysics, Volume ZI, Theory, Society
of Exploration Geophysicists, Tulsa, Oklahoma, pp. 658-08. c(2,2)=b(3);
Gustafsson, P., 1994. SPOT satellite data for exploration of fractured
aquifers in a semi-arid area in southeastern Botswana, Applied c(1*1)=b(6);
Hydrogeology, 2(2):9-18. ~(3,2)=b(2);
Henderson, R.G., and I. Zietz, 1967. The computation of second ver-
tical derivatives of geomagnetic fields, Mining Geophysics, Vol- c(2,3)=b(2);
ume ZI, Theory, Society of Exploration Geophysicists, Tulsa,
Oklahoma, pp. 606-20. c(l,3)=b(4):
Karnieli, A,, A. Meisels, L. Fisher, and Y.Arkin, 1993. Automatic c(3,1)=b(4);
extraction and evaluation of geological linear features from digi-
tal remote sensing data using a Hough transform, Proc. 9th Envi- c(1,2)=b(5);
ronmental Research Institute of Mchigan Thematic Conference
on Geologic Remote Sensing, Pasadena, California, pp. 299-310. c(2.l)=b(5);

PHOTOGRAMMETRICENGINEERING& REMOTE SENSING April 1999 457

 for j=l:3

for i 4 . 5

c(j,i)=cQ,&i);

end

end

for ii4:5

for j=l:5

c(ij)=c(6-i,j);

end

end

%%%% Which will crrale the fohwing outpa %%%%

mot =

2.4048 5.5201 8 6537 11.7915 14.9309 18.0711

rad~i=

0.0000 1.0000 14142 2.0000 2.2361 2.8284

A=

8.5254

PE&RS BACK ISSUES SALE IS your library incomplete?

A N Y SET O F 1 2 ISSUES ....................................... $ 7 5
For USA Addresses (postage included)
NO^-USA Addresses: ~ d $35
d for postage. Did someone b o r r o w an issue
DIRECTORY OF THE MAPPING SCIENCES ........$10 and not return it?
GISILIS ISSUE ........................................................ $ 1 0
ANY 1998 SPECIAL ISSUE ...................................$20 While supplies last, you can order
OTHER SINGLE ISSUES .......................................... $ 7 back issues o f PE&RS,
Add $3.00 postage per issue for Non-USA addresses.
GST is charged to residents of Canada only (GST #135193065).
Tax is calculated at 7% x(subtota1 + shipping charges).

Availability: 1993 - 1 998

Out of Print: January 1998, October & December 1997; June
1996; January 1994; March, July, August, September, &
October I993

TO ORDER, CONTACT:
ASPRS Distribution Center
PO Box 305
Annapolis Junction, MD 20701 -0305
tel: 301 -61 7-7812; fax: 301 -206-9789
e-mail: [email protected]
VISA, Mastercard, and American Express are accepted.

458 April 1999 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

View publication stats